How reliable is a reliability calculation?

Hanspeter Dinner, Director – Global Sales, KISSsoft AG, hanspeter.dinner@KISSsoft.com

Management summary
Main gearbox bearing failures are of great concern to operators of wind turbines and by extension to the gearbox
supplier. They have a strong interest in predicting this risk of failure or the reliability of bearings and the bearing
subsystem. The bearing subsystem reliability number for a required lifetime is the fraction of the total gearbox
population (provided it is a large population) that will survive the required lifetime without any bearing failing.

Figure 0-1 Typical wind turbine main gearbox, bearings in yellow.

This reliability number has obviously a major influence on the total cost of ownership of a gearbox and the prod-
uct in which it is used [23]. A higher, initial investment, e.g., in a safe design, will result in a higher reliability and
hence lower costs for repairs and from downtime. A lower initial investment may result in a lower reliability and
hence higher costs for e.g. spare parts. It is therefore of great interest to predict the reliability of a gearbox or its
bearing subsystem. Such predictions are based on the bearing life calculation and a model for the variation of the
component life, if a large number of components is considered.

Here, a particular effect contributing to the margin of error of such predictions is investigated. The reliability num-
ber from a reference calculation is compared to reliability numbers of calculations where different parameters or
assumptions used in the calculations are varied slightly. These variations occur e.g., if two parties (manufacturer
and user of the gearbox, manufacturer and certification body, two potential suppliers, …) perform such a calcula-
tion using the same bearing data and the same loads. It is found that even small variations have an effect in the
range of +-10 %-points on the calculated reliability number for required life, R(Hreq). This means that – when com-
paring two calculated reliability numbers – it is of utmost importance that the calculations are done with exactly
the same assumptions, calculation methods, loads, using the same tools and preferably by a single entity. This
basically means that a purchaser of gearboxes is advised against comparing two reliability numbers submitted by
two potential gearbox suppliers, but he should calculate these numbers by himself. It is too simple to “tune” such
calculated reliability numbers!

KISSsoft AG T. +41 55 254 20 50

A Gleason Company F. +41 55 254 20 51
Rosengartenstr. 4, 8608 Bubikon info@KISSsoft.com
Sharing Knowledge
Switzerland www.KISSsoft.com
Introduction
Objective of the study
Bearing failures are a major concern in wind turbine main gearboxes. Predicting the risk of a bearing failure (to
calculate the total cost of ownership TCO and to predict the need for spare parts) or a comparison between sev-
eral designs to find the one with the lowest risk (e.g. when selecting suppliers and their designs) is of interest to
all stakeholders.
A qualitative assessment of the risk of failure is achieved by calculating the bearing life L and therefrom the bear-
ing failure probability F(Hreq) for a required life Hreq. If this is done for all bearings in a gearbox, then, the bearing
subsystem reliability for the required life R(Hreq) may be calculated. The bearing subsystem reliability R(Hreq) may
then be used to calculate the bearing subsystem failure probability F(H req) = 1-R(Hreq). The bearing subsystem
failure probability for a required life, F(Hreq) is then the fraction of a gearbox fleet (assuming the fleet is a high
number of gearboxes) that will experience a failure of any bearing in the gearbox by design. This means it is an
“in-built” and expected fatigue failure.
The theory that is typically applied for such a calculation is well established and has been widely used e.g. in the
design process for e.g. aircraft gearboxes even in the 1970s, [11], [12], [13], [14], [15], [16], [17]. It is well docu-
mented, e.g. in [9] from where it may easily be implemented in a computer code. Commercial software is widely
available for system level reliability calculations as part of a design failure mode and effect analysis (DFMEA)
process, e.g. Reliasoft, [27] or within gear and gearbox design tools, e.g. KISSsoft [28]. It is nowadays applied to
wind gearboxes by turbine OEMs, operators, gearbox suppliers, [10] or academia, [25].

All stakeholders will be aware that a calculated reliability or rather risk of failure will not necessarily reflect the fail-
ure rate experienced in the field. Reason for this is for example the fact that only certain types of bearing failures
are open to a predictive reliability calculation. Basically, only the bearing rolling contact fatigue (surface or sub-
surface initiated) is available for a probabilistic assessment, see section 0 below. Also, tuning the calculations to
field observations to take into account e.g. site-specific loads is a necessity.

In this paper, another, simpler and therefore often overlooked aspect influencing the accuracy of such calcula-
tions is addressed. The calculated reliability as a function of time t, R(t) of a bearing subsystem (or any other sub-
system, e.g. the gears) will change with any small variation of input parameters or calculation method. Such vari-
ations may be due to different parties performing the calculations (this can be two engineers in the same com-
pany or two suppliers of gearboxes to one wind turbine manufacturer, or a gearbox supplier and a certification
agency) using different assumptions with respect to those input parameters. Practical experience in the industry
confirms this likely to be the case whenever more than one calculation is done, in particular, if the several calcu-
lations are done by different persons. The question then is: how much will the calculated reliability change if
somebody, by mistake or by intention, changes or “tunes” those input parameters? How reliable are such reliabil-
ity calculations?

Bearing failures in wind gearboxes

The 2014 wind turbine gearbox damage distribution based on NREL gearbox reliability database as shown be-
low, [23], is based on approximately 320 gearbox damage records. The majority of the damage occurs to bear-
ings (65 %), followed by gears (25 %).

09.08.2021 2 / 15
 Figure 0-1 Gearbox damage distribution, [23].

On the other hand, a quantitative assessment of the bearings is based on a static rating, typically along ISO 76
[6] a life rating along ISO 281 [5] and ISO/TS 16281 [7]. Furthermore, contact stresses may be compared to rec-
ommended levels given in industry specific standards and guidelines [1], [2], [3], [4]. However, these calculated
lifetimes and stress levels are covering only one or two (if we consider the static rating) possible failure modes of
bearing, that is, fatigue failure (due to surface or sub-surface-initiated fatigue). Other failure modes, in particular
white etching cracks or axial cracks, are not yet open to a quantitative assessment. Failure due to poor handling
or assembly technique is also not considered. This means that the calculations done during the design phase are
not addressing all potential future failure modes and therefore under predict the risk of failure or over predict the
reliability for a required lifetime (or overpredict the lifetime for a required reliability).

Figure 0-2 Bearing failure modes along ISO 15243 [24].

09.08.2021 3 / 15
Reliability as a design requirement for wind turbine gearboxes
ISO 81400-4, [1], states “… The required design life shall be specified for each of the major subsystems of the
gearbox including gears, bearings, housings, shafts and seals. To the extent possible, the specification should
identify the minimum required life, the reliability associated with the life calculation, and the method or standard to
be used for the calculation.”

In this statement, attention is directed to the underlying reliability of the material data or load capacity numbers
associated with the life calculation of components. For the bearing rating, applicable and typical design rules [1],
[2], [3], [4] and others usually stipulate the use of a bearing failure probability of 10 %. Wind turbine OEMs how-
ever often change this requirement to a lower failure probability of e.g. 5 %, giving an a1 factor a1 = 0.64 along
ISO 281, [5]. Correspondingly, the modified rating life is 0.64 times the basic rating life.

For gears, values for the permissible root and flank stress, Flim and Hlim respectively, are given for 1 % prob-
ability of damage, corresponding to a reliability level of 99 % for the gear rating against bending and pitting. How-
ever, ISO 6336 gear rating does include data for other levels of reliability. A more detailed approach is available
with AGMA 2101 where a reliability factor YZ is used. Said factor is equal to unity for one failure in 100 and in-
creases to YZ = 1.50 for fewer than one failure in 10’000. For shaft rating, DIN 743 again uses a different material
reliability level of 97.5 %.

This reliability requirement is valid only for the single component, or rather, the rating of a single component
against a single type of failure. E.g. if the gear rating is performed against bending with a reliability of 99 % and
the against pitting again with a reliability of 99 %, then, the reliability of the gear against bending or pitting is less
than 99 %.

Correspondingly, if in a gearbox each bearing individually just reaches the required life H req (e.g. 175’200 h = 20 y
as typically required in wind industry) at a probability of failure of F = 10% or a reliability of R = 90 %, then, the
reliability of the bearing subsystem for the required life is less than 90 %. The reliability decreases with the in-
creasing number z of bearings. To quantify the reduction, the reliability of the bearing subsystem (= all z bearings
together) R(Hreq) for a certain time Hreq is calculated from the reliability of each bearing individually, Rb(Hreq) as
follows:
z
R ( H req ) = Rb1 ( H req ) * Rb1 ( H req ) * ... * Rbz ( H req ) =  Rbi ( H req )
i =1

The underlying principle for above formula to be applicable is that the bearing subsystem is considered as a sys-
tem with a strictly serial structure. This means that if one bearing fails, there is no back-up and the bearing sub-
system is considered as failed. Other conditions are that the components and the subsystem is either “functional”
or “failed” and that the components are independent, their respective state “functional” or “failed” do not depend
on another components state.

Bearing reliability calculation

In 1939, W. Weibull developed a method and an equation for statistically evaluating the fracture strength of mate-
rials based upon small population sizes, [33]. This method can be and has been applied to analyze, determine,
and predict the cumulative statistical distribution of fatigue failure or any other phenomenon or physical charac-
teristic that manifests a statistical distribution. The dispersion in life for a group of homogeneous test specimens
can be expressed by:

  1   L − 
ln ln   =  * ln 
  R   L − 
  

09.08.2021 4 / 15
Where:
R = Probability of survival or reliability, 0 < R ≤ 1
 = Shape parameter or Weibull slope
L = life or number of cycles
 = location parameter or the time below which no failure occurs
L = Characteristic life, time at which 63.2 % of a population will fail (further below, T will be used for characteris-
tic life)

The format of above equation is referred to as a three parameter (3P) Weibull function. For most – if not all – fail-
ure phenomenon, there is a finite time period under operating conditions when no failure will occur. In other
words, there is zero probability of failure, or a 100 % probability of survival, for a period of time during which the
probability density function is non-negative. This value is represented by the location parameter . Without a sig-
nificantly large database, this value is difficult to determine with reasonable engineering or statistical certainty. As
a result,  was taken as zero by Lundberg and Palmgren and above equation was used in below, re-written form
for the prediction of bearing life vs. reliability, [33]:

  1   L 
ln ln   =  * ln 
  R  L 
  

This format is referred to as the two-parameter (2P) Weibull function. The estimated values of the Weibull slope 
and Lβ for the two-parameter Weibull analysis may not be equal to those of the three-parameter analysis, [12].
This concept is also the basis for the calculation of the a1 factor for the modified rating along ISO 281. Note that
the calculation for a1 factor has been changed in ISO 281:2007 (three parametric Weibull distribution is used with
a failure free time of  = 0.05*L; or C = 0.05) compared to ISO 281:1990 (where a two parametric Weibull distri-
bution is used and  = C = 0), [33]. In the latest revision of AGMA 6006, [3], for each bearing i, the reliability Rbi(t)
as a function of time t is calculated. Furthermore, for each bearing i, the reliability R bi(Hreq) for the required com-
ponent life Hreq is calculated. Note that the bearing life is taken as the modified reference rating life, L = L nmrh with
n = 10% or R0 = 90%. For the calculation of the bearing reliability, a three parametric Weibull distribution as
shown below is used:

 t − 
−  
  
R (t ) = e
L −
= 
− ln( R0 )

Where
L = calculated bearing life as modified reference rating Lnmrh
 = Location parameter = C * L; C=0.05
 = Shape parameter = 1.500
 = Scale parameter
R0 = Reference reliability, reliability used to calculate Lnmrh, R0 = 90%

The shape parameter  is in line with the value used in ISO 281, [5] and ISO / TR 1281, [31]. For the shape pa-
rameter , values in literature differ considerably, e.g. Harris [32], states “… for modern, ultraclean, vacuum re-
melted steels, values of e in the range 0.7-3.5 have been found…”.

The location parameter  = C*L = 0.05*L is also in line with above references. It is conservative when compared
with other values published, e.g.  = 0.053 in [30] or ftB = 0.1…0.3 in [9]. Note however that the failure free time
(or the location parameter) is defined in [31] with respect to the basic rating life L10 while in AGMA 6006, [3], the
calculated life L is the modified reference rating life L10mr.

09.08.2021 5 / 15
Influence of the shape of the Weibull distribution
AGMA 6006 uses a three parametric Weibull distribution for bearings and it defines the shape and location pa-
rameter to be used, see above. Let us compare different Weibull distributions according to relevant sources in-
cluding FAG Wälzlagerpraxis, [34], AGMA 6006 [3] (based on ISO 281:2007 [5]), Bertsche [9] (using parameters
as listed in section 7, mean values), ISO 281:1977 [35]. Below, the respective parameters are listed:

Type of Weibull distri- Shape parameter Location pa-

bution rameter
β γ
FAG 2 parametric 1.11 for ball, 1.125 for roller bearings 0*L
AGMA 6006 3 parametric 1.5 for ball, 1.5 for roller bearings 0.05*L
Bertsche 3 parametric 1.1 for ball, 1.35 for roller bearings 0.20*L
Table 0-1 Shape and location parameter for different bearing types according to different sources.

Using the above parameters for roller bearings, we find the below curves for reliability and failure probability over
time (where the time t is normalized by the characteristic time T). While all curves obviously yield identical results
for t = T, the results are quite different for a required reliability of R = 90% (horizontal solid line in red). There, we
find (for roller bearings) that the calculated life along FAG is t = 0.135*T (left vertical solid line in grey) whereas
along Bertsche, we find t = 0.355*T (right vertical dashed line in grey). Value for AGMA rating is t = 0.265*T (mid-
dle vertical dotted line in grey). Basically, we observe that for low lifetimes (as typically required in a design) the
choice of the parameters β (shape parameter) and γ (location parameter) is critical. A comparison between two
designs or two calculations for one design must therefore quite obviously use the same distribution (Weibull dis-
tribution) and the same parameters β and γ. Not discussed here is the influence if distribution functions other
than the Weibull distribution (e.g. normal distribution, log-normal distribution, …) although Thoma, [10], reports
that field data (of wind turbine gearboxes, considering gears and bearings failures) may be matched more closely
if a log-normal distribution is used.

Figure 0-1 Bearing reliability as a function of normalized time for different bearing types (shown for roller bearings only) and different
shape and location parameters along different sources (FAG, AGMA 6006 / ISO 281:2007, Bertsche and ISO 281:1997).

09.08.2021 6 / 15
Bearing subsystem life and reliability
Base line model and reference conditions
In this paper, the bearing subsystem of a 3MW class wind turbine main gearbox (MGB) as shown below is inves-
tigated. The gearbox uses four planets with four rows of CRB each in the low-speed stage LSS, three planets
with two rows of cylindrical roller bearings CRB in the intermediate speed stage ISS and a CRB and two taper
roller bearings TRB on the high-speed stage HSS shafts. Together with the planet carrier bearings, a total of 32
individual bearing rows are employed. For each bearing, the modified reference rating life L = L10nmrh is calculated
along ISO/TS 16281 using a KISSsoft, [28], model as shown below. Note that the gearbox is used with a main
shaft supported by two main bearings, hence, the loads on the planet carrier bearings are low and the corre-
sponding lifetimes very high.

Figure 0-1 Left: Bearing arrangement considered, note that HSS shaft is not shown, two CRB and a four point ball bearing FPBB is
used there, [21]. Right: KISSsoft model used for calculation of the modified reference rating life, the operating clearance, non-linear
bearing stiffness, load and stress distribution for all bearings. Planetary carrier ISS not shown, housing not shown.

For this paper, different effects will be considered in the calculation. From a reference condition, small changes
will be introduced to study how these changes affect the resulting reliability of the bearing subsystem. In the be
low table, the effects considered and their reference conditions are listed.

Property Reference value Affects Comments

Bearing clearance, posi-
tion within the tolerance
field
Shaft and housing / in-
ner and outer race tem-
perature
Load application posi-
tion with respect to gear
face width
Planet load distribution
(K)
Mean tolerance value
for bearing clearance,
shaft and housing toler-
ance
Temperature differences
are set as per Table 4 of
ISO 81400-4, [1]
Load is assumed to act
in center of gear face
width
K = 1.10 for LSS and
K = 1.05 is assumed for
bearing load calculation
Bearing operating clear-
ance and hence load
distribution in bearings
Bearing operating clear-
ance and hence load
distribution in bearings
Load distribution on
planet bearings LSS and
ISS
Load level on the planet
bearings LSS and ISS
The load distribution
then affects the rated life
The load distribution
then affects the rated life
No recommendations
are given in standards
or certification guide-
lines
K is a factor typically
used in gear rating but
uneven load distribution
among the planets also
affects bearing load

09.08.2021 7 / 15
 Nominal torque on gear-
box input
Lubricant temperature
Lubricant contamination
Bearing clearance varia-
tion in planet bearings
Pressure angle, gears
Table 0-1 Reference conditions for the calculation of the bearing life.
100% nominal load is
used
Selected as 65C° for all
bearings
Selected as - / 17 / 14,
beta25 = 75 along ISO
4406
All bearings have same
bearing clearance
Operating pressure an-
gle wt is used to calcu-
lated gear forces
Load level on all bear- To consider site specific
ings except carrier bear- loads
ings
Lubricant viscosity and
aISO factor
Life rating, through ec
value and aISO factor
Load distribution among Bearings assumed to be
the bearings, for LSS perfectly matched
and ISS planet bearings
Bearing forces Often, bearing supplier
uses nominal pressure
angle, not operating
pressure angle

Results, reference calculation

Using calculation settings in line with DNV GL guideline [4] and those listed in above table, the calculations per-
formed using KISSsoft software [28] resulted in the following individual bearing lifetime. Note that the modified
reference rating life in hours, L10mrh is calculated, this is in line with requirements of AGMA 6006 [3]. The life of the
planet carrier bearings is not a calculated but an assumed value, such that they will not noticeably influence the
reliability rating.

Stage Position Shaft L10mrh in hours

LSS RS-RS Planet, same result for all four planets 314'820
LSS RS-GS Planet, same result for all four planets 9'889'777
LSS GS-RS Planet, same result for all four planets 10'504'562
LSS GS-GS Planet, same result for all four planets 345'866
LSS RS Carrier 9'999'999
LSS GS Carrier 9'999'999
ISS RS Planet, same result for all two planets 646'878
ISS GS Planet, same result for all two planets 675'973
ISS RS Carrier 9'999'999
ISS GS Carrier 9'999'999
HSS RS Driving shaft 344'247
HSS GS-RS Driving shaft 9'999'999
HSS GS-GS Driving shaft 622'918
HSS RS Driven shaft 286'648
HSS GS-RS Driven shaft 677'429
HSS GS-GS Driven shaft 382'756
Table 0-1 Calculated bearing life, reference values. For carrier bearings, a very high life is assumed.

For each of the 32 bearings, the reliability function R(t) and the failure probability function F(t) is plotted using a
three parametric Weibull distribution with shape and location parameter along AGMA 6006,  = 1.50,  =
0.05*L10mrh. See grey lines in below figure. Then, the bearing subsystem reliability function (pink line in below

09.08.2021 8 / 15
figure) and the failure probability function (cyan line in below figure) is calculated as the product of all functions of
the individual curves.

The intersection of the line representing the required subsystem life Hreq at 175’200 hours (vertical, red) and the
line representing the time dependent subsystem reliability (blue) results in a subsystem reliability value of about
0.59 (blue, dashed, horizontal line). This means that in a large population of gearboxes, we expect that for 100%-
59%=41% of the gearboxes, a bearing fatigue failure is to be expected by design. If the required reliability of R req
= 90% should also apply for the bearing subsystem reliability (and not only for the individual bearing), then, the
rated life drops from 175’200 hours to about 72’000 hours (intersection of horizontal, solid cyan line and blue
curve, from there, vertical, solid pink line).

Figure 0-1 Single bearings reliability curves (grey), bearing subsystems reliability (pink), bearing subsystems failure probability (cyan),
required subsystems life H req (vertical, solid blue), rated reliability (horizontal, solid orange) and subsystem reliability at required life
(horizontal, dashed orange) and resulting subsystem life at rated reliability (vertical, dashed blue).

If we now investigate the rated reliability (the reliability for the required lifetime H req) by grouping the values per
stage (LSS planet bearings in orange, carrier bearings in blue, ISS planet bearings in pink, HSS bearings in
cyan), we observe in the below graphic that

▪ The outer planet bearings (RS-RS and GS-GS position) drive the reliability of the LSS
▪ The ISS has fairly high reliability values
▪ Bearing 30 on the output shaft has the lowest reliability

If we multiply the reliability values, we find in the second graphic below that the bearing subsystem reliability will
be determined by the LSS planet bearings reliability and to some extend by the HSS bearings. The results are in
line with field experience and expectations.

09.08.2021 9 / 15
Figure 0-2 Reliability for required life for all bearings, grouped per stage (blue: LSS planet bearings, green: LSS carrier bearings, grey:
ISS planet bearings, cyan: ISS carrier bearings, pink: HSS bearings).

Figure 0-3 Bearing group reliability per stage / group of bearings.

Variation of calculation settings

Parameters varied
Several experiments are set up. In each experiment, only one parameter is changed compared to the reference
calculation. This is a non-conservative approach.

09.08.2021 10 / 15
 Experi- Parameter varied Variation Comments
ment
1 Bearing clearance, position Lower position in tolerance Considers bearing clearance,
within the tolerance field field of resulting bearing clear- shaft and housing / planet
ance is used bore tolerance
2 Bearing clearance, position Upper position in tolerance Considers bearing clearance,
within the tolerance field field of resulting bearing clear- shaft and housing / planet
ance is used bore tolerance
3 Shaft and housing / inner and Temperature difference be- Influences bearing operating
outer race temperature tween races reduced by 5K clearance but not oil viscosity
4 Shaft and housing / inner and Temp. difference between Influences bearing operating
outer race temperature races increased by 5K clearance but not oil viscosity
5 Load application position with Offset of 1 cm from center
respect to gear face width
6 Load application position with Offset of 2 cm from center
respect to gear face width
7 Planet load distribution (K) K values increased by 0.05 in
both LSS and ISS
8 Planet load distribution (K) K values decreased by 0.05
in both LSS and ISS
9 Nominal torque on gearbox in- Nominal load decreased by To consider e.g. site-specific
put 2.5 % loads
10 Nominal torque on gearbox in- Nominal load increased by 2.5 To consider e.g. site-specific
put % loads
11 Lubricant contamination One class worse, - / 19 / 16, Compared to reference class -
beta25 = 7 / 17 / 14, beta25 = 75
11a Lubricant contamination One class better, -/15/12, Compared to reference class -
beta12=200 / 17 / 14, beta25 = 75
12 Pressure angle, gears Nominal instead of operating
pressure angle
13 Bearing clearance variation in Outer bearing rows, LSS plan- To simulate that bearings are
planet bearings ets, clearance reduced by 10 not perfectly matched
m, inner have a clearance in-
creased by 10 m
Table 0-1 Set up of the 14 experiments

Resulting reliability curves

For each experiment described above, the bearing lifetime for all 32 bearings were calculated (for the four planet
carrier bearings, a value of L = 9’999’999 h was used throughout). For each bearing in each experiment, the reli-
ability curve was calculated, and the bearing subsystem reliability and failure probability curve is plotted in the
below figure. This gives a cluster of 13 curves corresponding to the 13 experiments. The resulting reliability val-
ues for the required life Hreq are determined as intersection of the reliability curves with the vertical line at x = Hreq.
The resulting values are shown in the second figure below, again for all experiments. In both graphics, the high-
est (green), second highest (cyan), lowest (grey) and second lowest (magenta) curve are highlighted. Also, the
reference calculation result is shown in blue. Experiment 1 and experiment 12 gave highest reliability (same re-
sult was achieved when result is rounded to two digits), highlighted (green). Experiment 8 gave second highest
reliability (cyan). Experiment 2, experiment 6 and experiment 7 gave second lowest reliability (same result was
achieved when result is rounded to two digits), (pink). Experiment 11 gave lowest reliability (grey, solid line).

09.08.2021 11 / 15
Figure 0-1 Bearing subsystem reliability for different experiments (blue: reference calculation, green: highest result, pink: second low-
est result, grey: lowest result, cyan: second highest results, white: other results)

09.08.2021 12 / 15
Calculated reliability values R(Hreq) range from 38 % to 70 % or by 32 %-points. If we accept that experiment 11,
where a lower lubricant purity is used, is somewhat extreme (recommended purity levels to be used in calculation
are clearly documented e.g., in certification guidelines) and omit it from the assessment, we find as lower limit the
value of R(Hreq) = 51%, giving a range of 19%-points. The standard deviation for above values is σ = 8%-points if
highest and lowest values are considered. If experiment 11 is not considered, then, standard deviation reduces to
σ = 6(.4) %-points. Highest deviation from reference calculation is +11 %-points and -21 %-points (-8 %-points if
experiment 11 is not considered). Note that in the experiments, only a single parameter is varied compared to the
reference calculation while in a real-life scenario, it is likely that several parameters may deviate. From the
above, we clearly see that a +-10%-point error margin or more is to be expected if calculation settings are not
controlled with utmost care.

Conclusion
Governing effects
In line with prior experience, we find that the reliability of the bearing subsystem is governed by the outer bearing
rows in the LSS planet bearings and the output shaft bearings.
From above sensitivity study, we find that the most dramatic influence was observed if the lubricant cleanliness
level is changed. This is easy to understand as the lubricant cleanliness level itself affects the rated life of the
LSS planet bearings the most, since those have the lowest lubricant film thickness and therefore a low aISO fac-
tor. This means that for the LSS planet bearings, a low reliability at reference condition and a high variation of the
life and hence reliability if the lubricant cleanliness level changes are combined.
For the LSS planet bearings, other parameters are also difficult to assess. These include the load distribution
among the planets, in particular for systems with number of planets higher than three. Also, due to the high face
width of planets and due to system deformation, the gear forces typically do not act in the middle of the face
width, but the exact location is hardly known and changes over one rotation. Finally, since there are four bearing
rows in a single planet, there will be a variation in the bearing clearance even if they are matched. This again
leads to an uneven load distribution among the bearing rows.
On the HSS bearings, a major influence is the pre-tension of paired TRBs and the influence of the bearing race-
way temperature. On the output shaft bearings, additional loads due to generator misalignment will play a role
but is not investigated here. This is in line with experience that system alignment procedure, bearing assembly
and cooling is crucial for the reliability of the HSS bearings.
In this respect, the above calculations and the results gained are well in line with practical experience and under-
line the importance of sound bearing design, careful quality control in assembly to get the required clearances
and lubrication (in terms of temperature management and cleanliness).

How reliable is the reliability calculation?

From above analysis, we find that a calculated bearing subsystem reliability of a typical wind turbine gearbox has
a reproducibility error of +-10 %-points. This means that if two parties calculate the same gearbox, we must ex-
pect that their results will differ by the above error. This also means that if for two competing designs, the differ-
ence in the calculated reliability (assuming the calculations are done by the two respective suppliers) is not more
than 20 %-points, then, it is not clear whether the difference is accidental. The high level of reproducibility error
had to be expected as any life calculation is highly sensitive to the assumptions and process.

As a purchaser of a gearbox, interested in comparing the total cost of ownership for several competing designs,
based on gearbox reliability numbers supplied by the suppliers, this level of variation is disappointing. The obvi-
ous solution is that all calculations are done strictly with identical assumptions and methods with respect to all
parameters affecting them.

09.08.2021 13 / 15
References
[1] ISO 81400-4:2005, Wind turbines, Part 4, Design and specifications of gearboxes
[2] IEC 61400-4:2012, Wind turbines - Part 4: Design requirements for wind turbine gearboxes
[3] ANSI-AGMA 6006-B20, Standard for Design and Specifications of Gearboxes for Wind Turbines, revision of
ANSI/AGMA/AWEA 6006-A03
[4] Germanische Lloyd Rules and Guidelines Industrial Services, Guideline for the Certification of Wind Tur-
bines, 2010
[5] ISO 281:2007, Rolling Bearings – Dynamic Load Rating and Rating Life
[6] ISO 76:2006, Rolling Bearings – Static Load Ratings
[7] ISO/TR 16281:2008, Rolling Bearings – Methods for Calculating the Modified Reference Rating Life for Uni-
versally Loaded Bearings
[8] ISO 4406:1999, Hydraulic Fluid Power – Fluids – Method for Coding the Level of Contamination by Solid
Particles
[9] B. Bertsche, Reliability in Automotive and Mechanical Engineering, Springer
[10] F. Thoma, Determination of the reliability for a multi-mega-watt offshore gearbox, International Conference
on Gears 2015
[11] E. V. Zaretsky et al., Determination of Turboprop reduction Gearbox System Fatigue Life and Reliability,
NASA/TM 2007-215019
[12] E. V. Zaretsky, Rolling Bearing Life Prediction, Theory, and Application, NASA/TP 2013-215305
[13] E. V. Zaretsky et al., Weibull-Based Design Methodology for Rotating Structures in Aircraft Engines, Interna-
tional Journal of Rotating Machinery, 9: 313-325, 2003
[14] M. Savage et al., Reliability Model for Planetary Gear Trains, NASA Technical Memorandum 82859, 1982
[15] D. G. Lewicki et al., Fatigue life analysis of a Turboprop Reduction Gearbox, NASA Technical Memorandum
87014, 1985
[16] Detroit Diesel Allison, Study of turboprop systems reliability and maintenance costs, NASA CR 135192,
1978
[17] M. Savage et al., Computerized life and reliability modelling for turboprop transmissions, NASA Technical
Memorandum 100918, 1988
[18] S. S. Sheng, Report on wind turbine subsystem reliability – a survey of various databases, NREL/PR-5000-
59111, 2013
[19] S. S. Sheng et al., Wind turbine gearbox failure modes – a brief, ASME/STLE International Joint Tribology
Conference, 2011
[20] G. G. Antony, How to Determine the MTBF of gearboxes, powertransmissionengineering, April 2008
[21] Dinter et al., Epicyclic gearbox for a wind power installation, patent US 2010/0160104 A1, 2010
[22] ISO 6336-5, Calculation of load capacity of spur and helical gears – Part 5: Strength and quality of materials
[23] https://openei.org/datasets/dataset/2014-wind-turbine-gearbox-damage-distribution-based-on-the-nrel-gear-
box-reliability-database, 30.11.2017
[24] ISO 15243:2017, Rolling bearings – Damage and failures – Terms, characteristics and causes
[25] J. G. Calderon et al., Effects of Bearing Configuration in wind turbine gearbox reliability, Energy Procedia,
Vol. 80, 2015

09.08.2021 14 / 15
[26] J. Ukonsaari, Wind turbine gearboxes, maintenance effect on present and future gearboxes for wind tur-
bines, ISBN 978-91-7673-279-3
[27] www.reliasoft.com
[28] www.kisssoft.com
[29] Lundberg, G.; and Palmgren, A.: Dynamic Capacity of Rolling Bearings. Acta Polytech. Mech. Eng. Ser.,
vol.1, no. 3, 1947.
[30] Tallian, Tibor: Weibull Distribution of Rolling Contact Fatigue Life and Deviations Therefrom. ASLE
Trans.,vol. 5, no. 1, 1962, pp. 183–196.
[31] ISO/TR 1281-1:2008 Rolling bearings – Explanatory notes on ISO 281 – Part 1: Basic dynamic load rating
and basic rating life
[32] T. A. Harris, M. N. Kotzalas, Essential Concepts of Bearing Technology, 5th ed., CRC, 2007
[33] Hung Nguyen-Schäfer, Numerische Auslegung von Wälzlagern, Springer Vieweg, 2016
[34] FAG, Wälzlagerpraxis: Handbuch zur Gestaltung und Berechnung von Wälzlagerungen, 4. Auflage 2015
[35] ISO 281:1977Rolling bearings -- Dynamic load ratings and rating life

09.08.2021 15 / 15